Minecraft Diamond

MinecraftDiamondEnv is the Minecraft Diamond environment of DreamerV3 (Hafner et al., 2023), the benchmark on which DreamerV3 became the first agent to collect diamonds from scratch, without human data or curricula. It reproduces the reference implementation (embodied/envs/minecraft_flat.py), built on MineRL 0.4.4 and Minecraft 1.11.2.

For MineRL's own competition tasks (Treechop, Navigate, ObtainDiamond) with a generic discrete action set, use make_minecraft_env from simulusion.envs.minecraft_env instead.

Install

MineRL 0.4.4 runs a real Minecraft client. It needs Linux, Python 3.11, a JDK 8 and a display (or xvfb), and the patched wheel that adds the block break speed setting. The docker/envs image installs all of it, with DeepMind Lab and BSuite:

Terminal

docker build -f docker/envs/Dockerfile -t simulusion-envs .
docker run --gpus all --rm -v "$PWD/runs:/data/simulusion" simulusion-envs \
  simulusion train --config dreamer model.env_backend=minecraft model.env=diamond \
    model.action_repeat=1

Outside Docker, on Linux with Python 3.11:

Terminal

sudo apt-get install openjdk-8-jdk xvfb
pip install https://github.com/danijar/minerl/releases/download/v0.4.4-patched/minerl_mirror-0.4.4-cp311-cp311-linux_x86_64.whl
xvfb-run -a python your_script.py

The task


Goal	Obtain a diamond, through 12 milestone items
Rewards	+1 the first time per episode each of `log`, `planks`, `stick`, `crafting_table`, `wooden_pickaxe`, `cobblestone`, `stone_pickaxe`, `iron_ore`, `furnace`, `iron_ingot`, `iron_pickaxe`, `diamond` is held; plus 0.01 x the change in health (health scaled to [0, 1])
Episode end	Death, or 36000 steps (30 minutes at 20 Hz). Breaking diamond ore does not end the episode. If Minecraft's Malmo backend fails, the episode is cut short there (truncated, without reward) and the next reset restarts the game, as in the reference.
Block breaking	100x faster (`break_speed=100`), so the attack key need not be held
Actions	25: `noop`, `attack`, `turn_up/down/left/right` (15 degrees), `forward`, `back`, `left`, `right`, `jump` (with forward, held for 10 steps), `place_dirt`, and 13 abstract crafting, placing, equipping and smelting actions
Camera	Pitch kept within +-60 degrees

task="wood" is the reference's quick sanity task: the 12 basic actions and +1 per log collected.

The paper describes the health reward as +-0.01 per heart. The reference code, which produced the reported results, scales the health change by health in [0, 1], so a heart is worth 0.001; this environment follows the code.

Observations and actions

python

from simulusion import make_minecraft_diamond_env

env = make_minecraft_diamond_env("diamond", size=(64, 64))
obs = env.reset()
obs, reward, done, info = env.step(env.action_space.sample())
print(env.action_names[:4], info["achievements"]["log"])

Key	Shape	Content
`image`	`(3, 64, 64)` uint8	First-person view
`inventory`	`(num_items,)`	Count of every item
`inventory_max`	`(num_items,)`	Largest count this episode, which tells the agent which milestones it reached
`equipped`	`(num_types,)`	One-hot of the held item
`health`, `hunger`, `breath`	`(1,)`	Scaled to [0, 1]

Actions are one-hot vectors over env.action_names (an integer index works too). The action space is marked one_hot, so agents that choose their policy from the action space can use a categorical one. info["achievements"] reports which milestone items the episode has reached.

Dreamer config

python

from simulusion.configs.dreamer_config import DreamerConfig

cfg = DreamerConfig(
    env_backend="minecraft",
    env="diamond",
    action_repeat=1,
)

The agent acts on every Minecraft step, as in the reference, so action_repeat must be 1; make_env raises otherwise. Episodes last minecraft_length steps and time_limit does not apply.

Option	Default	Description
`cfg.env`	`"diamond"`	`"diamond"` or `"wood"`
`cfg.minecraft_break_speed`	`100.0`	Block breaking speed multiplier
`cfg.minecraft_length`	`36000`	Episode length in steps

Edit this page on GitHub

Updated Nov 4, 2026